Researchers at the University of Bristol have been working on a new gadget that might soon be able to offer assistance to those of us that do not like to stop and ask for directions. Their work builds upon the SCAMP architecture camera-processor chip that we previously reported on. The Pixel Processor Array (PPA) that makes up these chips contain a processor and memory for each individual pixel, and each processor is able to communicate with every other processor. This allows super fast, massively parallel computations to be performed directly on the image sensor.
The new work takes advantage of the advancements made in the SCAMP architecture to create a low-power device — that does not rely on external signals such as GPS — to tell you where you are currently located.
This image sensor-processor hybrid captures images of its surroundings. These images are heavily downsampled to either 8x8 or 8x4 pixels, depending on the method employed, to retain only key pieces of information. This information is then stored in a “database” of sorts — registers within the PPA are used to remember previously seen imagery. The image data is stored as a series of nodes that are spatially coherent such that, if needed for the application, it would be possible to associate GPS data with them.
The team then developed a custom algorithm that is used when the device is switched into its localization mode. In this mode, the sensor compares the current location being observed with all other images currently stored in the database, and returns the node in the internal topological map that is closest to the sensor’s present location. Each new image is able to be compared with every image in the database simultaneously.
The technique was tested against three databases of images to assess accuracy and performance. In each case, it was found to outperform the baseline analysis methods in terms of performance, while also running significantly faster. Image processing speed was clocked at over three hundred frames per second.
This work on PPAs may open the door to some interesting uses for low-power devices and robotics platforms in the future. It is not clear exactly how broadly applicable this particular method will be, however. Data storage on the focal plane of the PPA itself is necessarily limited, so it is unlikely that we will see a device based on this method that will recognize locations from a large geographic range any time soon, if ever.